Assessment and improvement of content-based image retrieval approaches for screening mammography analysis

Abstract

This study is trying to assess methods commonly used in content-based image retrieval (CBIR) for screening mammography analysis. A database consists of 20 different BI-RADS classes of mammogram patches taken from IRMA database is used in this study. Three feature extraction methods, namely grey-level co-occurrence matrix(GLCM), principal component analysis, and scale-invariant feature transform (SIFT) are being investigated through prior studies. Three retrieval methods are also studied, namely k-nearest neighbor (KNN), support vector machines (SVM), and mutual information (MI). From previous studies of those methods, we are trying to improve the CBIR approaches by combining the methods to see which one gives the better improvement. The result will be evaluated using precision and recall rate and areas under receiver operating characteristic (ROC) curve. The result of this study is expected to contribute more towards better Computed-Aided Diagnosis (CADx) and specifically screening mammography analysis in clinical cases.